Assessing and Enhancing the Robustness of Brain Tumor Segmentation using a Probabilistic Deep Learning Architecture

Motivation: Motivated by the challenge of enhancing the robustness of deep neural network decisions against variable noise in MRI-based brain tumor segmentation. Goal(s): This study aims to evaluate the efficacy of probabilistic bottlenecks in enhancing segmentation robustness. Approach: Our approach simulates structured perturbations at increasing strength to assess their impact on segmentation performance utilizing the Wasserstein distance between per-sample Dice score distributions and the sensitivity with respect to the perturbation strength. Results: Results show probabilistic bottlenecks significantly increase robustness to Gaussian noise, yet offer limited improvement towards Gaussian blur, with varying results for other perturbations, highlighting the perturbation-specific nature of network resilience. Impact: This study provides a tool to assess and guard against various perturbations in deep learning.It specifically demonstrates that probabilistic bottlenecks boost robustness of performance withrespect to certain noise types, but not all.